Your AI Cheat Sheet: Key Concepts in Common Sense Terms

Artificial intelligence (AI) is back, in a big way, and this time around it works. AI is being deployed across industries and in many problem areas. Some applications, like autonomous vehicles, appear on the front page of popular media; others are buried deep in the complex processes of enterprise systems. As AI technology continues to flow into the industry, it is important for everyone in the business of data analysis and decision-making to have a firm understanding of the technology’s core ideas and language. With that in mind, the following are some AI concepts explained in a clear language.

Model

A model is the end result of a machine-learning algorithm’s training process.

The basic function of an AI application is transformation of inputs into outputs. For example, give the application a picture of a cat and an AI application identifies (or predicts) it to be a cat. However, to get to the point where an AI application can predict, it needs to create a model of how inputs relate to outputs. This is completed through training.

During training, a machine-learning algorithm is fed a series of input/output matches or examples. By giving the algorithm many examples of inputs that match to the same output—different kinds of cats that are all cats—the algorithm is able to identify the core characteristics of a cat’s physical form. After being fed a variety of such inputs and outputs, the algorithm has developed an understanding of how different inputs relate to outputs. This understanding is called a “model.”

Artificial Intelligence

Defining AI is difficult largely because understanding human intelligence is difficult. A good chunk of the history of AI has been this introspective, philosophical process of self-discovery. After all, how can intelligent machines be built if humans cannot understand—let alone describe—intelligence. The study of psychology has allowed humans to measure and define, in general terms, intelligence, but it has never offered a thorough or deep enough explanation to allow a programmer to translate intelligence into some kind of working computer process.

So, as understanding of human intelligence has grown, so has clarity around the meaning and scope of “AI.”

A very practical definition of AI is a machine process that makes predictions or take actions based on data. Voice-activated assistants, product recommendations, email client optimization, etc.—these sorts of applications are all around and have become a daily part of life. While some people are concerned about the capabilities of these tools, they are in fact heavily constrained and really quite limited. The most sophisticated AI currently deployed only has a very narrow bandwidth in which it can be effective. It can play chess or go, but that is it.

The more idealistic definition of AI is a machine that has broken free of such limitations and has human-like, general intelligence. Such a system would be capable of solving across problem spaces. It could interpret Dostoyevski, help with a golf swing and pick the right assets for a portfolio. It is not known when this sort of AI will be possible, because humans have not yet achieved a rigorous definition of the goal: general intelligence. In the meantime, these more narrow AI systems that assist humans in decision-making can be observed and deployed.

Nature Language Processing

Natural language processing (NLP) is the field of AI that is dedicated to interfacing with human language. This is a particularly important application for AI, as so much of the world’s information is embedded in a medium ill-suited for computer systems—language. Natural language processing technology transforms text into data and unleashes machine scale processing on the now-accessible information, producing analytics, insights or decisions.

Machine Learning

Machine learning is an umbrella term used to describe two different processes. The first process is model creation, where an algorithm generates a model using training data—examples of correct decisions as determined by people. The second process is prediction, where the algorithm uses the model it produced in training to make predictions from new data. Machine learning is what makes it possible for machines to make decisions based on data, and it is the dividing line for what is referred to, practically speaking, when the term “AI” is used.

Supervised Learning

Supervised learning is the classic way machine-learning algorithms build their models during training. In this approach, a model is given a dataset composed of input/output matches (annotated data) and asked to infer the connection.

Unsupervised Learning

Unsupervised learning is also about algorithms interpreting data to produce models, but there is a twist—the data the machine trains on are not annotated by people. Instead of being given a set of input/output matches, the algorithm is left to infer relationships from raw data. An NLP task known as “document clustering” is a good example of this approach. In document clustering, a machine-learning algorithm is given a variety of documents and asked to group them (or cluster them) based on patterns it identifies on its own.

Deep Learning and Neural Networks

Deep learning is a branch of machine learning that uses neural network technology to tackle more difficult tasks than classic machine learning allowed for.

The outputs a machine-learning algorithm generates are only as accurate as the models they run, and many tasks, like long-form translation, exceed the capacity of traditional models. Classic algorithms could not produce sufficiently precise models; they could not learn enough from the data.

Taking inspiration from the design of our human brains, neural net architecture is a vast set of layers, each capturing a specific detail of the input/output relationship. This approach allows for machines to build much more nuanced models—or learn more deeply—during training and, therefore, handle tasks previously thought impossible for computer systems.

Summary

Hopefully, the constellation of concepts that is AI has come into better focus, and comfort with the relevant terms is growing. Machine learning should no longer be such a mystery. Neural nets should seem (at least somewhat) less magical. While only some of the space’s major terms have been reviewed, these basics will be a useful foundation for exploring what is likely the most important technology of the 21st century.

Steve Cohen, COO, Basis Technology, Cambridge, MA, USA, stevec@basistech.com

Leave a Reply